Qwen3-8B-SFT-evmSmith-w0-642

Qwen3-8B-SFT-evmSmith-w0-642

This is the longer W0 LoRA adapter for Qwen/Qwen3-8B, trained as a warm-start checkpoint for the Forest of Audits / OPD-Env smart-contract audit-agent pipeline.

It is the same staged W0 recipe as the earlier 221-step adapter, trained longer:

• Process stage: 532 global steps.
• Answer stage: 110 additional steps.
• Final global step: 642.
• Dataset: pranay5255/forest-of-audits-w0-sft-qwen3-8b.
• Dataset revision: b5587bc1a5f1f9f52134f5525b1ee06ae0b2722c.
• Base model: Qwen/Qwen3-8B.
• LoRA rank: 16.
• LoRA alpha: 32.
• LoRA dropout: 0.05.
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, down_proj, up_proj.

Intended Use

This adapter is a research artifact for comparing W0 static gold-audit SFT checkpoints before moving to agentic EVMBench detect-mode evaluation. It is not intended for production security review or live smart-contract audit advice.

Evaluation

Single-turn smoke/focused evals are logged to the existing Trackio Space: pranay5255/qwen3-w0-trackio.

The paired baseline checkpoint is: pranay5255/Qwen3-8B-SFT-evmSmith at global step 221.